How Enterprises Govern AI Agents Across Multiple Models

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: How Enterprises Govern AI Agents Across Multiple Models · Channel: Databricks · 31:33 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, AI governance and regulation, AI safety

Summary

The video explores how AI agents are evolving from single-turn assistants into fleets that can carry out business processes, and what enterprises need to deploy them reliably at scale. Databricks presents Unity AI Gateway as a centralized governance layer for using multiple models and tools, with access controls, cost management, observability, and auditability. The discussion covers how coding agents are changing software development, the security challenges of personal agents accessing sensitive data, and layered approaches to monitoring and permissions. Speakers also describe the emerging agentic stack, the value of open-source harnesses, and expanding use cases for both technical and non-technical users.

Key points

Enterprises are moving from narrow, single-turn agents toward fleets of agents that can handle end-to-end business processes.
Databricks frames its approach around choice of models and tools, business context, and governance and control.
Unity AI Gateway provides a centralized governance layer for multiple models and tools, including access policies, guardrails, budgets, alerts, caps, observability, and audit trails.
Coding agents are evolving from autocomplete and code authoring into systems that can delegate work, coordinate sessions, and manage ongoing operations.
Developers continue to contribute through taste, creativity, defining what should be built, and making decisions with better context.
Personal and ephemeral software can be created quickly, but agents that access sensitive business data need sandboxing, restricted permissions, and monitoring.
Speakers recommend layered defenses, including network controls, sandboxing, permission limits, gateways, and agent-based reviews of activity.
A useful enterprise agent stack connects agents to the right business data and tools, while supporting orchestration, permissions, and deployment.
Open-source harnesses can give organizations flexibility and confidence when building on agent infrastructure.
Agent use is expanding beyond software engineering into everyday assistance, infrastructure management, business questions, and other organizational tasks.

Chapters

Start	Chapter	Summary
00:01	From agent experiments to enterprise applications	The introduction explains the shift from single-turn agents to business-process applications and presents choice, context, and control as the key operational needs.
04:10	How agentic development is changing	The speakers discuss the rapid progression from chat and autocomplete to agents that write software, delegate tasks, and coordinate ongoing work.
11:25	The changing role of developers	The conversation focuses on how developers use judgment and creativity, while prototypes and agents make it easier to explore ideas and clarify requirements.
17:35	Governing personal agents and sensitive data	The speakers examine how enterprises can secure personal software and agents that run continuously or access valuable business data.
21:09	Layered security and agent oversight	The discussion advocates monitoring agents with intelligence and combining safeguards such as sandboxing, gateways, permission limits, and automated reviews.
24:17	The agentic stack and open-source harnesses	The speakers describe an emerging stack built around access to enterprise data, agent harnesses, orchestration, plugins, and the flexibility of open source.
28:56	Use cases beyond coding	The closing section highlights agents as general-purpose assistants for tasks across work and life, including use by non-technical employees.
31:02	Enterprise applications and closing remarks	The speakers note that customers are building agentic external applications and conclude by emphasizing the pace of change.

Quotes

It’s not just the building part, but really operationalizing Agentic apps in production. — 01:03

taste creativity like you still need to know what you want — 11:25

you govern intelligence with intelligence. — 21:09

you need a Swiss cheese kind of approach. — 22:43

The big shift is that it’s no longer a coding agent. — 29:29

Concepts and entities

Unity AI Gateway, Databricks, Patrick Wendell, Peter Steinberger, OpenClaw, OpenAI, Tibo Sato, Codex, Stellantis, Harness, Unity AI Gateway, defesa em profundidade

Full transcript (original language)

00:01 Hi everyone, welcome and thank you for joining us. Today we’re going to talk about Agentic applications and what it takes to move from experiments to build, deploy, orchestrate and govern them at scale. So what are Agentic apps? Across nearly every industry, teams have started with single turn agents for things like AI assistance for specific and and and somewhat narrow tasks. And now they’re moving towards building and deploying fleets of agents. So these are full

00:31 agentic applications that can handle complete business processes from customer service to claims processing and more. And this is where things get interesting. Um, it gets complicated because you’ve built once and you created the agent loop. But then each agent has to connect to models, harnesses, and other agents and then the business data that sits in databases and external systems. And then once you deploy, you’re managing access and monitoring performance. And you don’t want to have to rebuild the scaffolding

01:03 each time you get to an agent in production. So, while building Agentic apps is getting easier, the real question is how you move from one-off builds to something much more repeatable. And that’s really the shift we’re focused on today. It’s not just the building part, but really operationalizing Agentic apps in production. At Data Bricks, we make this happen through choice, context, and control. And so choice is really the freedom to

01:33 use the best model or agent or harness MCP or the tool for the job. Context is really building agents that know your business data and they know the context and that’s really the way that they’re most useful um right away. And then the third area is control and this is providing governance, cost controls and observability. AI governance in particular has been top of mind for our customers. You know, your teams are building and they’re using a lot of AI more than ever before, but the governance and the visibility

02:03 into AI usage and spend has just not kept up. And that’s why today we’re excited to announce the general availability of Unity AI gateway. This is a major milestone in our vision for AI governance on data bricks. And with Unity AI gateway, you get a single governed AI control plane. So what does this mean in practice? So Unity AI gateway gives you one governance layer for multiAI. So this means that you can use the best model or

02:34 agent harness for each task. These are natively hosted frontier models through to the latest open source models and frameworks and you can use whatever comes next without rebuilding controls for every provider or tool. Unity AI gateway helps you control costs. You have access policies, guardrails, spend management tools, things like budgets, alerts, and caps. And it provides end toend observability into which models and agents are used, which tools they call and where latency and cost are coming from. And this

03:04 really helps you optimize and ensure continuous improvement. With general availability, you can now use centralized governance for foundational models and MCP servers for Google Drive, GitHub, SharePoint, Atlassian, Slack, and all the tools that you typically use for business. Plus, you get built-in cost controls and full lineage and audit trails. And this really helps you move from experimentation to production scale AI. So we have an exciting agenda lined up for you today. We’ll start with a Q&A featuring Patrick Wendell, co-founder of

03:36 data bricks and leader of our AI engineering team. And he’s joined by Peter Steinberger, the creator of OpenClaw and a member of the technical staff at OpenAI, as well as Tibo Sato, who leads Codeex development at OpenAI. Then Amber Roberts will show you how to build and deploy an Agentic apps on data bricks. And we’ll hear from Hugo Sicher. He’s the head of Agentic AI at Stalantis, a leading global manufacturer behind brands you know like Jeep, Chrysler, and Dodge. So a lot of big topics to cover. Let’s dive in.

04:10 So first let’s talk about the evolution of building. With the explosion of coding agents, agent harnesses, orchestration frameworks, the way developers are building has fundamentally changed. Tibon, maybe we can start with you. How have you seen development evolve and what’s the new paradigm? >> It’s evolved very quickly. I feel um one silver lining to me is like humans adapt very very fast. Uh like even like a year year and a half ago, I remember

04:41 like people always elided. I was just like I can ask Chachi PT about like a coding question you know maybe I copy paste it then autocomplete became very efficient as well like um folks were like spending time in their IDEs and getting better and better autocompletes and it was like autocompleting entire files at some point and then now we’re like in the aentic era where you’re just sort of like delegating to an agent that does like most of the work and then goes much more beyond just the programming as as well like just basically like just

05:12 goes and and understands a lot of like the requirements, a lot of the world around it. Um, and then it just like thinks through the entire thing like just very much like a human does like when you delegate to another human. And then now we’re like you know into like exploring an entirely new realm which is like you know entire factories where you know you have multiple agents automatically like also delegate to each other and like you’re plugging them in the right systems and then now you’re not actually as as as the user you know necessarily delegating the things. There’s like very very fast evolution

05:42 from you know the very beginnings of like you know having like an intelligent thing a chatbot that you can just ask questions to like now it’s like okay it’s like a lot of software is being written automatically and you know maybe you don’t even review all of the code anymore. >> We had a year ago we had the CLI and you had to be careful to clear the session um and like you made multiple multiple terminals. Well, like over the last 6 months, so much changed. I would say, yeah, a year,

06:13 maybe even. Um, now you don’t you don’t have to care about compaction anymore, right? It like I used to I used to be the biggest CLI fan. When they they told me about the app, I like I don’t care. [laughter] And now like I when I go back to the seal, I’m like, how did I work with this, you know? Now, now like I I don’t think in in prompts anymore. like I just have my pinned topics or like tasks and then like I keep I keep asking question updating it. Um

06:46 when I work on something bigger I use the session where I work in as my orchestrator. I just tell it, hey, create work trees and then like use codex to to pile it up into tasks and then like this session can orchestrate like 10 other codec sessions and then and then watch what they’re doing like pile the work together and then maybe create two, three, four, five, six PRs. So, so like it really got agents got so good that they now can be your orchestrator at the same time. Is that a

07:17 loop? Yeah. [laughter] But yeah, it kind of started like it was like, oh, it accelerates me a little bit now. I don’t have to search or like, you know, I get like a qualitative, you know, answer for like, you know, part of a task that I’m trying to and now it feels like, you know, you’re putting this jetack on and you’re you’re going like 10,000 miles an hour and you’re doing things you were never able to do before. So, it’s Yeah. Change very quickly. >> Also, why even type anymore, right? Now you have I mean we used to have voice already but it it would still be it will still convert to text and then I press

07:49 enter and now I use TPD voice and I just talk to it and it can orchestrate my sessions and it could do it in line. If it’s more complex it like does it in the sidebar. Um and I still think we’re early. It’s still it’s still too much like like text. >> Mhm. Maybe I can quickly say at data bricks um our answer to to that question which is really similar to a lot of the things that you all are talking and obviously we’re big CEX uh fans so we’re often using the Codex app um or the the Codex

08:20 and the Chachupt app but um I think for for us the biggest change in the the recent past is moving beyond just code authoring as what people are doing with these agentic tools to actually having you know in the data platform space for instance just keeping data pipelines up and running and dealing with operational issues and observing them and seeing oh my gosh we had a cost overrun here or this data is late. Um I I would call those kind of observability and ops type challenges and a lot of the work we’re

08:51 doing now at data bricks is is unleashing agents to do that type of work which is actually sort of different than hey I’m building a code asset right now and I’m adding a particular feature. It’s this horizontal kind of like neverending ops type uh type work and I think that’s for us a frontier that we’re seeing is just the ability of these agents to have very longunning horizontal responsibilities. I I was actually curious to ask you guys if you see that as well inside of OpenAI as you look at more than just like let’s say authoring a code asset but doing something over a long period. It’s it’s

09:22 it’s kind of surreal at opening eyes like there’s not a single meeting or event or crisis or anything where you know codeex or agents are not involved. It’s like you you have a data question is you know it’s going to be codeex is going to go and look at you know all the tables needed bricks and like the full lake and then you know like process it all combined with like some information that it has about the business how it works like the semantics of it all and then synthesizes all of that um in like a report and it does that in like five minutes and you know that’s how we do

09:54 that that’s how all business questions that I have about um you know openi like get answers just like codeex um and then like when we have SESS it’s like you know well Codex goes and like solves it most of the time um or you know when we do deploys and we want to for example we deployed uh the new voice in the TJ app it’s like it was not it was not like an openi employee that went and you know entered like 10% in stats sig and then 20% it’s like it was just like codex rolled it out over the course of a day

10:24 um and that I think that changed like super super quickly and now we’re so used to it um and then we’re also entering like I think another layer of delegation where you like you watch this and then it’s like another agent is actually watching the other agent and then now suddenly you’re like automating a bigger and bigger chunk of your organization over time. >> I mean I use this all the time, right? I have like one long running session that monitors a workflow that have deployed on GitHub just to see like is it efficient? are there any problems? And

10:54 like it wakes up every 10 minutes and checks on it and like it I I I used GitHub as a database and like stored all the results and it it it yesterday it warned me that that’s now 30 GB and and this and this part of the subtask is now slow and recommended me like as resolution. So you just have and that was just a prompt you know like it it the session learned as as it went along and as as I steered it but it’s it’s really just the session that’s not monitoring my deployment. So Peter with with all of this automation this agentic

11:25 automation what’s left for the developer what do you think is the role of the developer >> taste creativity like you still need to know what you want like I mean you can ask a model but like that’s not going to usually not going to be what’s going to be most successful >> I am curious like do you view it as you know there’s some amount of the job that’s like specification I think that’s where taste is coming like what what ought to be done here um then there’s there’s the actuals of like hey you know h maybe how are the internals being

11:56 built is it working is it not working where do you think how far are we in your guys view of like developers are just merely in the business of specification at this point uh versus like um you know they’re still having to really understand intricate details about architecture and so on so forth >> I think that was always like the natural like um if if you look at software engineers in like large tech companies. It’s like the evolution of um you know

12:26 what you expect from someone in like a normal career evolution. It’s like you know that you start you’re an IC hands on keyboard you’re coding every day and then like more and more you’re like you know maybe you know you’re working with other um engineers that started bit more junior you tell them you know like how to work with the codebase and you know you coach them and then suddenly you know you find yourself like being a tech lead and you know you’re writing docs and um you know you’re planning things a

12:56 little bit more and then you know you evolve into like maybe like the TLM role you know where like the majority of your times just like talking to people and writing docs. Um, [laughter] >> but also >> I think that’s like a natural >> also as an engineer we always built automations. We always built things that like help us get more stuff done faster. And now just like like thinking what’s the best way to use all this intelligence that’s that’s just automation. It’s it’s just now that we have like more powerful tools and like a single person can do

13:27 even more. It’s like you have this idea and you’re like, “Oh, it’s like I would have never done this because this, you know, would have taken me two days, but you know, now it’s just like, you know, it’s like, yeah, like do it.” >> Like if if someone someone asks me, “Hey, wouldn’t this feature be cool?” I I kind of get mad because like [laughter] like why wouldn’t you just build build a prototype in Codex? >> Like it doesn’t matter. Like, okay, maybe it runs for two days and it’s like it’s it’s unsealed, so it might not be exactly what you want, but you mean you can like try it out. And then like and then asked me, hey, wouldn’t that be

13:58 cool? and I can click around and I immediately have a much better idea how big how big of an impact would have on the code base how could that look like uh you know we have to in this new and age the the most important thing we have is attention so with all these better tools we can do much more preparation ahead so that when when when I have to think about it I have more context and I can make [clears throat] better decisions faster. >> I think this is yeah it resonates quite a bit. I mean at data bricks no one is

14:29 more empowered than our product managers these days because you know the old way would be like okay maybe you make a visual a very static visual mockup of an idea and you try to convince you know some engineering team to to look at what you’re interested in building but now as you’ve said like just the standard way of proposing ideas is to go and build a pretty featurerich prototype from engineers PMs or or anyone in the company and for me the main thing the main takeaway for me has been that I’ve realized how underspecified in the p the

15:00 past way of doing it was you know like um just because at best we could maybe give a few workflows or give some static images or some but there’s so much richness that’s lost when you attempt to describe an idea in that way compared to like just delivering a prototype and I mean prototype is even maybe underelling it a lot of these are like fairly functional even even being delivered by by codecs over the period of a few hours um it’s just like it’s a whole new way of basically even deciding what to build

15:31 and it it’s obviously so much better than it was in the past is my experience. you you know that the best way is like sometimes when you have meetings I’ll I’ll press my my my record button while we talk and then like Codex gets all this rambling and like just start building and that you couldn’t do this six months ago but now the the models are are so good at converting this full bar stuff into into like the intent the intent that comes out. Um and many times if you just if if you just

16:02 give them more and more even even if parts of the composition are completely unrelated um it’ll get you the better result. >> Yeah, we we’ve seen this with models over time like now it’s like just really you you’re very surprised that you just like ramble at it and it’s like okay you actually understood the intent. It’s like it’s like three months ago that would not have been possible. >> Yeah. >> But back to your point is like I think there are two kinds of software that’s like emerging. It’s like a personal software and also ephemeral uh software where like you’re not sharing like a static mockup anymore. You’re like just

16:34 quickly getting a full prototype going and then you share it. Um or you’re just building software just for yourself which you don’t actually want to share. It’s just like it’s doing like one task. Maybe it’s doing it for a couple of days or like a week or a month. Maybe you maintain it, maybe you don’t. but it’s like almost free and there’s like no desire to actually like make this into like a share tool or infrastructure and maintain it for others. It’s like not a very personal thing or it’s ephemeral just to prove a point or exchange like you know get an idea through and this is something that was definitely not

17:05 possible before and like we’re seeing um with sites uh TGD sites is like you know you could just ask uh codeex or tagged work to like create a site and then share it and a lot of like the prototypes are being shared like this at openai >> it’s great like I I used to you know a year ago maybe you got a briefing via email and now like I get a I get an HTML page like a really well designed. This is everything you need to the customer and it’s kind of like it’s it’s really nice to to read. So like yeah, it makes every day every day is better through

17:35 that >> when you have everyone experimenting like this and they’re moving really fast which is great. How do you think about governing all of this first at the developer level and then as they’re deploying it in production at the org level? >> Yeah. Yeah. Maybe I can I can talk about what we’ve seen and I would love to hear from the OpenAI crew as well. Like we absolutely see this pattern TBO you mentioned of of people have kind of their own personal they can build such highly customized software which is which is really amazing because everyone wants to work in slightly different ways or slightly different um you know they

18:05 have slightly different problems they’re trying to solve but one thing one challenge we had early on that we’ve been thinking a lot about and I know open eye is looking at things in this area is like okay so let’s say you have 10,000 employees that’s how many we have at data bricks and they want to deploy their own little personal software components And some of it can run on their personal devices, but some of it they want to, you know, be running all the time. They want it to run and and we frankly want it to run somewhere secure where we’re not, you know, those that software by definition should be touching fairly valuable data if it’s

18:35 going to be something kind of useful. Um, so this is a, you know, we we’ve built different mechanisms internally, some of which are also in our our products around, you know, can you sandbox this stuff? Can you observe its behavior? Can you downscope its permissions and so forth? Um but I’m curious if you’ve you’ve looked at that or or how you guys think about this this um when that software is touching very sensitive data basically. >> Yeah. Right. Right now sites don’t support um reading data dynamically. So you package it all as part of a site.

19:06 It’s like fully sandboxed. Uh it’s restricted in access by default. It’s only the creator who gets access and then you know you have like the full arback controls that you’re used to. you could share with a team or you can share with other individuals. Um and because what agents and modern like you know five six souls like they’re like so capable they can read so much data and like remangle it and rejigle it and like re it’s like you don’t really know where that data came from like you know what’s in the asset. So you have to be very

19:37 careful about how you manage access. So that’s like by default it’s like fully sandbox access is like very restricted and then you know users get to decide if they want to share a little bit more. >> Yeah. Nice. I I would say we’re just we’re just starting in this area but one one thing we’ve also built is like anytime there’s access from one of these personal pieces of software to a a business system it gets mediated through a proxy. Basically it’s we call it our our gateway AI gateway. But um that way we’re able to kind of you know OOTH gives you so much in terms of okay you

20:07 can you can say okay the average person should maybe they can have read access but not write access their agents can have maybe read access but not right access to critical system so they can’t go delete a bunch of stuff if there’s a mistake but even then you know we do want to observe things like huh suddenly you know this personal agent read read 15,000 documents in a period of an hour that’s not you know normal behavior. So um so it’s kind of become this uh this interesting security and governance question how you treat these little I would call them personal agentic pieces

20:38 of software because they’re not quite they don’t quite represent the security model is not quite right to treat them as that person because they’re maybe slightly less trusted than the person because the person may have pure intent but they accidentally you know uh um encoded something improper in their agent but they’re also not really uh similar to like third party proper enterprise apps, you know. So, I think that’s been a a a thing we’ve been thinking a lot about is how to once you automate a lot of useful workflows in your company and maybe sensitive data is being touched, how do you handle that?

21:09 And I know um Peter, I know you’ve been also thinking about things along these lines. >> I I mean the only way you can do this is like you you you govern intelligence with intelligence. So, so if you you look at network patterns, if there’s anything that’s unusual is being flagged, then maybe a human looks over it. I mean, I have my cases where like people ping me and ask me, “What is this doing?” And I was like, “Oh, what is this actually doing?” >> Yeah, I agree. I think there needs to be this kind of constant monitoring and intelligence layer just checking are the

21:39 agents doing things that seem reasonable or maybe seem less reasonable and maybe, you know, throwing up flags or blocking behavior and or at least asking questions if something odd is going on. >> When we released Codex originally had two modes. I had the full sandbox mode which asks you for permission for everything and then the like run you know it’s like what we also called like yolo it’s like run fully outside of the sandbox which we don’t recommend full file system access full internet access no human supervision no supervision at all um and with models becoming more and

22:11 more capable this is also like increasing the risk in general like you know maybe you know by mistake it does something or you know it gets interested in something that you know it shouldn’t read and it’s like you know and suddenly it’s like doing something and it’s going off the rails. Um so it’s very important with capabilities is increasing that you know you have this like this other layer of intelligence and another agent that just monitors the primary agent and there are like multiple ways of doing this but we built this auto review mode um which is essentially a full-blown agent going over the full trajectory all

22:43 the tool calls and you know making sure that uh no unnecessary risk is being taken and then to your point I think also like the single choke hold that the gateways are becoming where you have all the data flowing through. They’re an excellent way to achieve this as well. >> Yeah. And it seems to me like you need a Swiss cheese kind of approach. Like no one of these mechanisms is going to be bulletproof. But if you layer the mechanisms like maybe you have, you know, you just have some basic network firewalling and sandboxing. So you you can’t trivially just send data out and you maybe have like a little supervisor

23:14 locally looking at what tool calls are being made. And then maybe also have a gateway that’s doing some uh correlation analysis across agents. and you have downscoping of permissions. So you you know no one of these is perfect but I think if you combine these I feel like you’ll end up with something fairly robust. >> Having defense and depth right is like always the right approach I think. >> Yeah. I mean humans can make mistakes as well. So you you just need to make sure to have enough layers. Like even that auto review thing I I was like when I heard that this is the new default open I’m

23:45 like oh it’s going to be so annoying. But now I’m I’m kind of I’m kind of glad because because it’s like an additional layer. If I would do something where I’m not thinking, the system would like go into my head and ask me, “Are you really sure?” >> Yeah. Yeah. And I think they’re evolving to be far better. The usability of some of the early, you know, it would be like, “Hey, are you okay running this bash command, Patrick?” And it would give me a bash command that was, you know, 100. >> It gives you a bash command that’s like this long. Yeah. >> I’m like, I have no hope of even understanding that bash command, let alone like scrutinizing its security

24:17 properties. So, I feel like we’re um we’re advancing pretty quickly towards far better techniques than that. >> So, I want to flip to a different topic, slightly different topic, and that’s really operating at scale. So, we talked a lot about builders and moving quickly and building agents and a lot of customers are now thinking about if I’m trying to scale, I don’t want to have to rebuild a scaffolding each time, right? What is I think it’s a hot question right now. What is the agentic stack, right? What do you need to really be able to build at scale? um wondering if your perspective on that. Maybe Tibo, we

24:48 can start with you and then go around to the others as well. >> There’s different levels of investment. Um I think often just the biggest bang for buck is just like give everyone access to an agent that is like well configured has access to the right data within the enterprise. um like if you don’t connect it to any data source like for example like my favorite Slack channel is like the one where we have this agent connect to data bricks and it’s just like an ask any question but

25:19 it also has access to uh the company’s like you know business and a couple of like you know documents and it also has access to sack so it’s just like sort of has this run well-rounded information and I can ask it very sophisticated questions and if it didn’t have access to any of this data obviously it’s not going to be very useful um and so giving everyone access to agent that they can configure and that the company has configured. Super important. And then you enter like levels of automation where you want to, you know, run a harness. It’s like it’s

25:49 just like running, you know, the back end like, you know, just like on your infrastructure and it’s like not even managed by, you know, like specific humans or not prompted by specific humans that are many different ways like open clause is awesome at that. We’re seeing like very successful enterprise deployment. We also have like the codeex harness which is also fully open source. But we’re doing a lot of open source things. Um and in general it’s like it’s goes over like a period of um experimentation I would say even within enterprise with routers in front of it gateways um you know different ways of

26:20 like getting access to plugins like plugins becoming like an open spec that we’re participating in. Um so getting access to context being able to give the permissions to to take action and um it’s still very much being uh developed but I think you know that’s sort of like the stack that’s emerging. Yeah, I think the harnesses I mean just as like a a user basically the harnesses are just so powerful that they’re almost becoming a part of our like internal applications doc at data bricks almost and and this is uh this is where the fact that um

26:51 codeex is open source is is like really awesome because you know it goes back to if you’re betting on a core piece of infrastructure you kind of want to know that it’s always going to be there. you know, we’re making many multi-year bets and we might even be building parts of a production engineering stack on uh on a a on a harness like this. Maybe you wouldn’t use the word harness. I don’t know if I if I’m using the appropriate language, but like um but uh the fact that it’s open source and and liberally licensed for us has been huge at data bricks because it’s just let us know that okay it’s not going to be a thing

27:22 where like one day the license on this thing changes and like our whole company has to stop if uh if if there’s some you license change or renegotiation or something like that. So so I think that’s been a great area in which you’ve been a real leader. Um obviously open call from the beginning uh um if I understand Peter was open source I think is that right? And and then uh >> I think that was the first commit I did was the MIT license. >> Yeah exactly. So like I think you guys have been real leaders in that and like as someone building on top that’s that’s a huge deal for us at data bricks.

27:55 >> Yeah. And then it’s like it it’s what I find fascinating. It’s like you have the code of the agent itself, the harness and then and then you also have the agent and then you know it’s like if you want to fork it and change it it’s like you know you can just go and do it. It’s all there. You can make it your own. Um we’re seeing a lot of successful forks and deployments. >> That’s where we have our best ideas from like SLGO came from I think Froex. >> That’s an internal fork. >> Yeah. Yeah, I was going to say, is that

28:26 a a person’s name? Because like what are the odds? Um, >> no, no, no. I mean, I have I have Claudex like where where I play around with ideas because not every idea makes you into the mainline product, right? You kind of >> it’s like sometimes you think this be cool and then actually you don’t use it and you don’t want to create a product that has like a thousand features that are that are not well thought through. So the best way to to like figure it out is to try it and it’s just very easy to like play around with it like Codex changes Codex and then they can like tweak.

28:56 >> So So we talked about how people are building. Can you tell me a little bit about what you’re seeing as the top use cases like where are customers being successful? >> I mean the big shift is that it’s no longer a coding agent. It’s now people realize how much more they can do. Like it it helps me in in every every step of my life. Like I automate something where I get an email every morning with like these are the most important things. Like I use it for infrastructure management. Of course I I do coding but also sometimes I just like I ask you to

29:29 to like explain me something and then like you can visualize me stuff and I can like make it interactive. It’s it’s like the the universal encyclopedia almost globe. >> It’s [laughter] Exactly. Um yeah, like when we started with with Codex and coding agents, it was all about coding and then it’s like oh the models are like much more general and you can do anything with it. It’s like if you connect it to the right uh data sources through plugins and with catch work like what we saw it’s like it’s

30:00 really taken off and we now have more non-technical users of the same agent. Like under the hood catch work is the same agent. It’s like the same harness as codecs, same capabilities, same plugins. Um, but we just packaged it in a way which is like much more intelligible and also safe for like, you know, day-to-day use by everyone. And so we now have more users of this technology that are non-technical than technical. Um, we’re going to keep pushing that. But I think, you know, it’s just it’s so empowering for everyone, you know, across organizations

30:32 so that you just get like this little personal assistant in your pocket. It works on mobile, works on web, works in the desktop app. You can pretty much ask it like anything. think you can be super ambitious and then often more often than not you’re just kind of surprised just like oh it can actually do that. That’s why those I mean I I’ll keep calling it codeex but like that’s why that the name change made sense because like I basically use it like other people use the chat GBT for and now people are discovering this little tab between chat and work and like work is gives you like a actually powerful machine with like

31:02 codex under the hood where if you have all your connectors set up there there’s your there’s your remote machine where you can run anything. >> Yeah. Cool. Well, I mean, I’ll just end by saying maybe um we’ve seen very similar. I think the internal productivity is obviously like a massive use case and highly facilitated by these models. Um but you know, we’re also just seeing even external applications being built by our customers that are themselves agentic and some of them may even use you know harnesses underneath and they’re they’re moving to that even from like their own custom agent loops

31:33 inside of their apps. So, it’s a it’s a really exciting time uh and stuff is moving so fast and and uh uh thank you to both of you guys for joining us. We really appreciate it. >> Thank you. Thanks for having us.